1 //===- LoopSimplify.cpp - Loop Canonicalization Pass ----------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This pass performs several transformations to transform natural loops into a 11 // simpler form, which makes subsequent analyses and transformations simpler and 12 // more effective. 13 // 14 // Loop pre-header insertion guarantees that there is a single, non-critical 15 // entry edge from outside of the loop to the loop header. This simplifies a 16 // number of analyses and transformations, such as LICM. 17 // 18 // Loop exit-block insertion guarantees that all exit blocks from the loop 19 // (blocks which are outside of the loop that have predecessors inside of the 20 // loop) only have predecessors from inside of the loop (and are thus dominated 21 // by the loop header). This simplifies transformations such as store-sinking 22 // that are built into LICM. 23 // 24 // This pass also guarantees that loops will have exactly one backedge. 25 // 26 // Indirectbr instructions introduce several complications. If the loop 27 // contains or is entered by an indirectbr instruction, it may not be possible 28 // to transform the loop and make these guarantees. Client code should check 29 // that these conditions are true before relying on them. 30 // 31 // Note that the simplifycfg pass will clean up blocks which are split out but 32 // end up being unnecessary, so usage of this pass should not pessimize 33 // generated code. 34 // 35 // This pass obviously modifies the CFG, but updates loop information and 36 // dominator information. 37 // 38 //===----------------------------------------------------------------------===// 39 40 #define DEBUG_TYPE "loop-simplify" 41 #include "llvm/Transforms/Scalar.h" 42 #include "llvm/ADT/DepthFirstIterator.h" 43 #include "llvm/ADT/SetOperations.h" 44 #include "llvm/ADT/SetVector.h" 45 #include "llvm/ADT/Statistic.h" 46 #include "llvm/Analysis/AliasAnalysis.h" 47 #include "llvm/Analysis/DependenceAnalysis.h" 48 #include "llvm/Analysis/Dominators.h" 49 #include "llvm/Analysis/InstructionSimplify.h" 50 #include "llvm/Analysis/LoopPass.h" 51 #include "llvm/Analysis/ScalarEvolution.h" 52 #include "llvm/IR/Constants.h" 53 #include "llvm/IR/Function.h" 54 #include "llvm/IR/Instructions.h" 55 #include "llvm/IR/IntrinsicInst.h" 56 #include "llvm/IR/LLVMContext.h" 57 #include "llvm/IR/Type.h" 58 #include "llvm/Support/CFG.h" 59 #include "llvm/Support/Debug.h" 60 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 61 #include "llvm/Transforms/Utils/Local.h" 62 using namespace llvm; 63 64 STATISTIC(NumInserted, "Number of pre-header or exit blocks inserted"); 65 STATISTIC(NumNested , "Number of nested loops split out"); 66 67 namespace { 68 struct LoopSimplify : public LoopPass { 69 static char ID; // Pass identification, replacement for typeid 70 LoopSimplify() : LoopPass(ID) { 71 initializeLoopSimplifyPass(*PassRegistry::getPassRegistry()); 72 } 73 74 // AA - If we have an alias analysis object to update, this is it, otherwise 75 // this is null. 76 AliasAnalysis *AA; 77 LoopInfo *LI; 78 DominatorTree *DT; 79 ScalarEvolution *SE; 80 Loop *L; 81 virtual bool runOnLoop(Loop *L, LPPassManager &LPM); 82 83 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 84 // We need loop information to identify the loops... 85 AU.addRequired<DominatorTree>(); 86 AU.addPreserved<DominatorTree>(); 87 88 AU.addRequired<LoopInfo>(); 89 AU.addPreserved<LoopInfo>(); 90 91 AU.addPreserved<AliasAnalysis>(); 92 AU.addPreserved<ScalarEvolution>(); 93 AU.addPreserved<DependenceAnalysis>(); 94 AU.addPreservedID(BreakCriticalEdgesID); // No critical edges added. 95 } 96 97 /// verifyAnalysis() - Verify LoopSimplifyForm's guarantees. 98 void verifyAnalysis() const; 99 100 private: 101 bool ProcessLoop(Loop *L, LPPassManager &LPM); 102 BasicBlock *RewriteLoopExitBlock(Loop *L, BasicBlock *Exit); 103 BasicBlock *InsertPreheaderForLoop(Loop *L); 104 Loop *SeparateNestedLoop(Loop *L, LPPassManager &LPM, 105 BasicBlock *Preheader); 106 BasicBlock *InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader); 107 void PlaceSplitBlockCarefully(BasicBlock *NewBB, 108 SmallVectorImpl<BasicBlock*> &SplitPreds, 109 Loop *L); 110 }; 111 } 112 113 char LoopSimplify::ID = 0; 114 INITIALIZE_PASS_BEGIN(LoopSimplify, "loop-simplify", 115 "Canonicalize natural loops", true, false) 116 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 117 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 118 INITIALIZE_PASS_END(LoopSimplify, "loop-simplify", 119 "Canonicalize natural loops", true, false) 120 121 // Publicly exposed interface to pass... 122 char &llvm::LoopSimplifyID = LoopSimplify::ID; 123 Pass *llvm::createLoopSimplifyPass() { return new LoopSimplify(); } 124 125 /// runOnLoop - Run down all loops in the CFG (recursively, but we could do 126 /// it in any convenient order) inserting preheaders... 127 /// 128 bool LoopSimplify::runOnLoop(Loop *l, LPPassManager &LPM) { 129 L = l; 130 bool Changed = false; 131 LI = &getAnalysis<LoopInfo>(); 132 AA = getAnalysisIfAvailable<AliasAnalysis>(); 133 DT = &getAnalysis<DominatorTree>(); 134 SE = getAnalysisIfAvailable<ScalarEvolution>(); 135 136 Changed |= ProcessLoop(L, LPM); 137 138 return Changed; 139 } 140 141 /// ProcessLoop - Walk the loop structure in depth first order, ensuring that 142 /// all loops have preheaders. 143 /// 144 bool LoopSimplify::ProcessLoop(Loop *L, LPPassManager &LPM) { 145 bool Changed = false; 146 ReprocessLoop: 147 148 // Check to see that no blocks (other than the header) in this loop have 149 // predecessors that are not in the loop. This is not valid for natural 150 // loops, but can occur if the blocks are unreachable. Since they are 151 // unreachable we can just shamelessly delete those CFG edges! 152 for (Loop::block_iterator BB = L->block_begin(), E = L->block_end(); 153 BB != E; ++BB) { 154 if (*BB == L->getHeader()) continue; 155 156 SmallPtrSet<BasicBlock*, 4> BadPreds; 157 for (pred_iterator PI = pred_begin(*BB), 158 PE = pred_end(*BB); PI != PE; ++PI) { 159 BasicBlock *P = *PI; 160 if (!L->contains(P)) 161 BadPreds.insert(P); 162 } 163 164 // Delete each unique out-of-loop (and thus dead) predecessor. 165 for (SmallPtrSet<BasicBlock*, 4>::iterator I = BadPreds.begin(), 166 E = BadPreds.end(); I != E; ++I) { 167 168 DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor " 169 << (*I)->getName() << "\n"); 170 171 // Inform each successor of each dead pred. 172 for (succ_iterator SI = succ_begin(*I), SE = succ_end(*I); SI != SE; ++SI) 173 (*SI)->removePredecessor(*I); 174 // Zap the dead pred's terminator and replace it with unreachable. 175 TerminatorInst *TI = (*I)->getTerminator(); 176 TI->replaceAllUsesWith(UndefValue::get(TI->getType())); 177 (*I)->getTerminator()->eraseFromParent(); 178 new UnreachableInst((*I)->getContext(), *I); 179 Changed = true; 180 } 181 } 182 183 // If there are exiting blocks with branches on undef, resolve the undef in 184 // the direction which will exit the loop. This will help simplify loop 185 // trip count computations. 186 SmallVector<BasicBlock*, 8> ExitingBlocks; 187 L->getExitingBlocks(ExitingBlocks); 188 for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(), 189 E = ExitingBlocks.end(); I != E; ++I) 190 if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator())) 191 if (BI->isConditional()) { 192 if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) { 193 194 DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in " 195 << (*I)->getName() << "\n"); 196 197 BI->setCondition(ConstantInt::get(Cond->getType(), 198 !L->contains(BI->getSuccessor(0)))); 199 200 // This may make the loop analyzable, force SCEV recomputation. 201 if (SE) 202 SE->forgetLoop(L); 203 204 Changed = true; 205 } 206 } 207 208 // Does the loop already have a preheader? If so, don't insert one. 209 BasicBlock *Preheader = L->getLoopPreheader(); 210 if (!Preheader) { 211 Preheader = InsertPreheaderForLoop(L); 212 if (Preheader) { 213 ++NumInserted; 214 Changed = true; 215 } 216 } 217 218 // Next, check to make sure that all exit nodes of the loop only have 219 // predecessors that are inside of the loop. This check guarantees that the 220 // loop preheader/header will dominate the exit blocks. If the exit block has 221 // predecessors from outside of the loop, split the edge now. 222 SmallVector<BasicBlock*, 8> ExitBlocks; 223 L->getExitBlocks(ExitBlocks); 224 225 SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(), 226 ExitBlocks.end()); 227 for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(), 228 E = ExitBlockSet.end(); I != E; ++I) { 229 BasicBlock *ExitBlock = *I; 230 for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock); 231 PI != PE; ++PI) 232 // Must be exactly this loop: no subloops, parent loops, or non-loop preds 233 // allowed. 234 if (!L->contains(*PI)) { 235 if (RewriteLoopExitBlock(L, ExitBlock)) { 236 ++NumInserted; 237 Changed = true; 238 } 239 break; 240 } 241 } 242 243 // If the header has more than two predecessors at this point (from the 244 // preheader and from multiple backedges), we must adjust the loop. 245 BasicBlock *LoopLatch = L->getLoopLatch(); 246 if (!LoopLatch) { 247 // If this is really a nested loop, rip it out into a child loop. Don't do 248 // this for loops with a giant number of backedges, just factor them into a 249 // common backedge instead. 250 if (L->getNumBackEdges() < 8) { 251 if (SeparateNestedLoop(L, LPM, Preheader)) { 252 ++NumNested; 253 // This is a big restructuring change, reprocess the whole loop. 254 Changed = true; 255 // GCC doesn't tail recursion eliminate this. 256 goto ReprocessLoop; 257 } 258 } 259 260 // If we either couldn't, or didn't want to, identify nesting of the loops, 261 // insert a new block that all backedges target, then make it jump to the 262 // loop header. 263 LoopLatch = InsertUniqueBackedgeBlock(L, Preheader); 264 if (LoopLatch) { 265 ++NumInserted; 266 Changed = true; 267 } 268 } 269 270 // Scan over the PHI nodes in the loop header. Since they now have only two 271 // incoming values (the loop is canonicalized), we may have simplified the PHI 272 // down to 'X = phi [X, Y]', which should be replaced with 'Y'. 273 PHINode *PN; 274 for (BasicBlock::iterator I = L->getHeader()->begin(); 275 (PN = dyn_cast<PHINode>(I++)); ) 276 if (Value *V = SimplifyInstruction(PN, 0, 0, DT)) { 277 if (AA) AA->deleteValue(PN); 278 if (SE) SE->forgetValue(PN); 279 PN->replaceAllUsesWith(V); 280 PN->eraseFromParent(); 281 } 282 283 // If this loop has multiple exits and the exits all go to the same 284 // block, attempt to merge the exits. This helps several passes, such 285 // as LoopRotation, which do not support loops with multiple exits. 286 // SimplifyCFG also does this (and this code uses the same utility 287 // function), however this code is loop-aware, where SimplifyCFG is 288 // not. That gives it the advantage of being able to hoist 289 // loop-invariant instructions out of the way to open up more 290 // opportunities, and the disadvantage of having the responsibility 291 // to preserve dominator information. 292 bool UniqueExit = true; 293 if (!ExitBlocks.empty()) 294 for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i) 295 if (ExitBlocks[i] != ExitBlocks[0]) { 296 UniqueExit = false; 297 break; 298 } 299 if (UniqueExit) { 300 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 301 BasicBlock *ExitingBlock = ExitingBlocks[i]; 302 if (!ExitingBlock->getSinglePredecessor()) continue; 303 BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 304 if (!BI || !BI->isConditional()) continue; 305 CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition()); 306 if (!CI || CI->getParent() != ExitingBlock) continue; 307 308 // Attempt to hoist out all instructions except for the 309 // comparison and the branch. 310 bool AllInvariant = true; 311 for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) { 312 Instruction *Inst = I++; 313 // Skip debug info intrinsics. 314 if (isa<DbgInfoIntrinsic>(Inst)) 315 continue; 316 if (Inst == CI) 317 continue; 318 if (!L->makeLoopInvariant(Inst, Changed, 319 Preheader ? Preheader->getTerminator() : 0)) { 320 AllInvariant = false; 321 break; 322 } 323 } 324 if (!AllInvariant) continue; 325 326 // The block has now been cleared of all instructions except for 327 // a comparison and a conditional branch. SimplifyCFG may be able 328 // to fold it now. 329 if (!FoldBranchToCommonDest(BI)) continue; 330 331 // Success. The block is now dead, so remove it from the loop, 332 // update the dominator tree and delete it. 333 DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block " 334 << ExitingBlock->getName() << "\n"); 335 336 // If any reachable control flow within this loop has changed, notify 337 // ScalarEvolution. Currently assume the parent loop doesn't change 338 // (spliting edges doesn't count). If blocks, CFG edges, or other values 339 // in the parent loop change, then we need call to forgetLoop() for the 340 // parent instead. 341 if (SE) 342 SE->forgetLoop(L); 343 344 assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock)); 345 Changed = true; 346 LI->removeBlock(ExitingBlock); 347 348 DomTreeNode *Node = DT->getNode(ExitingBlock); 349 const std::vector<DomTreeNodeBase<BasicBlock> *> &Children = 350 Node->getChildren(); 351 while (!Children.empty()) { 352 DomTreeNode *Child = Children.front(); 353 DT->changeImmediateDominator(Child, Node->getIDom()); 354 } 355 DT->eraseNode(ExitingBlock); 356 357 BI->getSuccessor(0)->removePredecessor(ExitingBlock); 358 BI->getSuccessor(1)->removePredecessor(ExitingBlock); 359 ExitingBlock->eraseFromParent(); 360 } 361 } 362 363 return Changed; 364 } 365 366 /// InsertPreheaderForLoop - Once we discover that a loop doesn't have a 367 /// preheader, this method is called to insert one. This method has two phases: 368 /// preheader insertion and analysis updating. 369 /// 370 BasicBlock *LoopSimplify::InsertPreheaderForLoop(Loop *L) { 371 BasicBlock *Header = L->getHeader(); 372 373 // Compute the set of predecessors of the loop that are not in the loop. 374 SmallVector<BasicBlock*, 8> OutsideBlocks; 375 for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header); 376 PI != PE; ++PI) { 377 BasicBlock *P = *PI; 378 if (!L->contains(P)) { // Coming in from outside the loop? 379 // If the loop is branched to from an indirect branch, we won't 380 // be able to fully transform the loop, because it prohibits 381 // edge splitting. 382 if (isa<IndirectBrInst>(P->getTerminator())) return 0; 383 384 // Keep track of it. 385 OutsideBlocks.push_back(P); 386 } 387 } 388 389 // Split out the loop pre-header. 390 BasicBlock *PreheaderBB; 391 if (!Header->isLandingPad()) { 392 PreheaderBB = SplitBlockPredecessors(Header, OutsideBlocks, ".preheader", 393 this); 394 } else { 395 SmallVector<BasicBlock*, 2> NewBBs; 396 SplitLandingPadPredecessors(Header, OutsideBlocks, ".preheader", 397 ".split-lp", this, NewBBs); 398 PreheaderBB = NewBBs[0]; 399 } 400 401 PreheaderBB->getTerminator()->setDebugLoc( 402 Header->getFirstNonPHI()->getDebugLoc()); 403 DEBUG(dbgs() << "LoopSimplify: Creating pre-header " 404 << PreheaderBB->getName() << "\n"); 405 406 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 407 // code layout too horribly. 408 PlaceSplitBlockCarefully(PreheaderBB, OutsideBlocks, L); 409 410 return PreheaderBB; 411 } 412 413 /// RewriteLoopExitBlock - Ensure that the loop preheader dominates all exit 414 /// blocks. This method is used to split exit blocks that have predecessors 415 /// outside of the loop. 416 BasicBlock *LoopSimplify::RewriteLoopExitBlock(Loop *L, BasicBlock *Exit) { 417 SmallVector<BasicBlock*, 8> LoopBlocks; 418 for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) { 419 BasicBlock *P = *I; 420 if (L->contains(P)) { 421 // Don't do this if the loop is exited via an indirect branch. 422 if (isa<IndirectBrInst>(P->getTerminator())) return 0; 423 424 LoopBlocks.push_back(P); 425 } 426 } 427 428 assert(!LoopBlocks.empty() && "No edges coming in from outside the loop?"); 429 BasicBlock *NewExitBB = 0; 430 431 if (Exit->isLandingPad()) { 432 SmallVector<BasicBlock*, 2> NewBBs; 433 SplitLandingPadPredecessors(Exit, ArrayRef<BasicBlock*>(&LoopBlocks[0], 434 LoopBlocks.size()), 435 ".loopexit", ".nonloopexit", 436 this, NewBBs); 437 NewExitBB = NewBBs[0]; 438 } else { 439 NewExitBB = SplitBlockPredecessors(Exit, LoopBlocks, ".loopexit", this); 440 } 441 442 DEBUG(dbgs() << "LoopSimplify: Creating dedicated exit block " 443 << NewExitBB->getName() << "\n"); 444 return NewExitBB; 445 } 446 447 /// AddBlockAndPredsToSet - Add the specified block, and all of its 448 /// predecessors, to the specified set, if it's not already in there. Stop 449 /// predecessor traversal when we reach StopBlock. 450 static void AddBlockAndPredsToSet(BasicBlock *InputBB, BasicBlock *StopBlock, 451 std::set<BasicBlock*> &Blocks) { 452 std::vector<BasicBlock *> WorkList; 453 WorkList.push_back(InputBB); 454 do { 455 BasicBlock *BB = WorkList.back(); WorkList.pop_back(); 456 if (Blocks.insert(BB).second && BB != StopBlock) 457 // If BB is not already processed and it is not a stop block then 458 // insert its predecessor in the work list 459 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) { 460 BasicBlock *WBB = *I; 461 WorkList.push_back(WBB); 462 } 463 } while(!WorkList.empty()); 464 } 465 466 /// FindPHIToPartitionLoops - The first part of loop-nestification is to find a 467 /// PHI node that tells us how to partition the loops. 468 static PHINode *FindPHIToPartitionLoops(Loop *L, DominatorTree *DT, 469 AliasAnalysis *AA, LoopInfo *LI) { 470 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ) { 471 PHINode *PN = cast<PHINode>(I); 472 ++I; 473 if (Value *V = SimplifyInstruction(PN, 0, 0, DT)) { 474 // This is a degenerate PHI already, don't modify it! 475 PN->replaceAllUsesWith(V); 476 if (AA) AA->deleteValue(PN); 477 PN->eraseFromParent(); 478 continue; 479 } 480 481 // Scan this PHI node looking for a use of the PHI node by itself. 482 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 483 if (PN->getIncomingValue(i) == PN && 484 L->contains(PN->getIncomingBlock(i))) 485 // We found something tasty to remove. 486 return PN; 487 } 488 return 0; 489 } 490 491 // PlaceSplitBlockCarefully - If the block isn't already, move the new block to 492 // right after some 'outside block' block. This prevents the preheader from 493 // being placed inside the loop body, e.g. when the loop hasn't been rotated. 494 void LoopSimplify::PlaceSplitBlockCarefully(BasicBlock *NewBB, 495 SmallVectorImpl<BasicBlock*> &SplitPreds, 496 Loop *L) { 497 // Check to see if NewBB is already well placed. 498 Function::iterator BBI = NewBB; --BBI; 499 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 500 if (&*BBI == SplitPreds[i]) 501 return; 502 } 503 504 // If it isn't already after an outside block, move it after one. This is 505 // always good as it makes the uncond branch from the outside block into a 506 // fall-through. 507 508 // Figure out *which* outside block to put this after. Prefer an outside 509 // block that neighbors a BB actually in the loop. 510 BasicBlock *FoundBB = 0; 511 for (unsigned i = 0, e = SplitPreds.size(); i != e; ++i) { 512 Function::iterator BBI = SplitPreds[i]; 513 if (++BBI != NewBB->getParent()->end() && 514 L->contains(BBI)) { 515 FoundBB = SplitPreds[i]; 516 break; 517 } 518 } 519 520 // If our heuristic for a *good* bb to place this after doesn't find 521 // anything, just pick something. It's likely better than leaving it within 522 // the loop. 523 if (!FoundBB) 524 FoundBB = SplitPreds[0]; 525 NewBB->moveAfter(FoundBB); 526 } 527 528 529 /// SeparateNestedLoop - If this loop has multiple backedges, try to pull one of 530 /// them out into a nested loop. This is important for code that looks like 531 /// this: 532 /// 533 /// Loop: 534 /// ... 535 /// br cond, Loop, Next 536 /// ... 537 /// br cond2, Loop, Out 538 /// 539 /// To identify this common case, we look at the PHI nodes in the header of the 540 /// loop. PHI nodes with unchanging values on one backedge correspond to values 541 /// that change in the "outer" loop, but not in the "inner" loop. 542 /// 543 /// If we are able to separate out a loop, return the new outer loop that was 544 /// created. 545 /// 546 Loop *LoopSimplify::SeparateNestedLoop(Loop *L, LPPassManager &LPM, 547 BasicBlock *Preheader) { 548 // Don't try to separate loops without a preheader. 549 if (!Preheader) 550 return 0; 551 552 // The header is not a landing pad; preheader insertion should ensure this. 553 assert(!L->getHeader()->isLandingPad() && 554 "Can't insert backedge to landing pad"); 555 556 PHINode *PN = FindPHIToPartitionLoops(L, DT, AA, LI); 557 if (PN == 0) return 0; // No known way to partition. 558 559 // Pull out all predecessors that have varying values in the loop. This 560 // handles the case when a PHI node has multiple instances of itself as 561 // arguments. 562 SmallVector<BasicBlock*, 8> OuterLoopPreds; 563 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 564 if (PN->getIncomingValue(i) != PN || 565 !L->contains(PN->getIncomingBlock(i))) { 566 // We can't split indirectbr edges. 567 if (isa<IndirectBrInst>(PN->getIncomingBlock(i)->getTerminator())) 568 return 0; 569 OuterLoopPreds.push_back(PN->getIncomingBlock(i)); 570 } 571 } 572 DEBUG(dbgs() << "LoopSimplify: Splitting out a new outer loop\n"); 573 574 // If ScalarEvolution is around and knows anything about values in 575 // this loop, tell it to forget them, because we're about to 576 // substantially change it. 577 if (SE) 578 SE->forgetLoop(L); 579 580 BasicBlock *Header = L->getHeader(); 581 BasicBlock *NewBB = 582 SplitBlockPredecessors(Header, OuterLoopPreds, ".outer", this); 583 584 // Make sure that NewBB is put someplace intelligent, which doesn't mess up 585 // code layout too horribly. 586 PlaceSplitBlockCarefully(NewBB, OuterLoopPreds, L); 587 588 // Create the new outer loop. 589 Loop *NewOuter = new Loop(); 590 591 // Change the parent loop to use the outer loop as its child now. 592 if (Loop *Parent = L->getParentLoop()) 593 Parent->replaceChildLoopWith(L, NewOuter); 594 else 595 LI->changeTopLevelLoop(L, NewOuter); 596 597 // L is now a subloop of our outer loop. 598 NewOuter->addChildLoop(L); 599 600 // Add the new loop to the pass manager queue. 601 LPM.insertLoopIntoQueue(NewOuter); 602 603 for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); 604 I != E; ++I) 605 NewOuter->addBlockEntry(*I); 606 607 // Now reset the header in L, which had been moved by 608 // SplitBlockPredecessors for the outer loop. 609 L->moveToHeader(Header); 610 611 // Determine which blocks should stay in L and which should be moved out to 612 // the Outer loop now. 613 std::set<BasicBlock*> BlocksInL; 614 for (pred_iterator PI=pred_begin(Header), E = pred_end(Header); PI!=E; ++PI) { 615 BasicBlock *P = *PI; 616 if (DT->dominates(Header, P)) 617 AddBlockAndPredsToSet(P, Header, BlocksInL); 618 } 619 620 // Scan all of the loop children of L, moving them to OuterLoop if they are 621 // not part of the inner loop. 622 const std::vector<Loop*> &SubLoops = L->getSubLoops(); 623 for (size_t I = 0; I != SubLoops.size(); ) 624 if (BlocksInL.count(SubLoops[I]->getHeader())) 625 ++I; // Loop remains in L 626 else 627 NewOuter->addChildLoop(L->removeChildLoop(SubLoops.begin() + I)); 628 629 // Now that we know which blocks are in L and which need to be moved to 630 // OuterLoop, move any blocks that need it. 631 for (unsigned i = 0; i != L->getBlocks().size(); ++i) { 632 BasicBlock *BB = L->getBlocks()[i]; 633 if (!BlocksInL.count(BB)) { 634 // Move this block to the parent, updating the exit blocks sets 635 L->removeBlockFromLoop(BB); 636 if ((*LI)[BB] == L) 637 LI->changeLoopFor(BB, NewOuter); 638 --i; 639 } 640 } 641 642 return NewOuter; 643 } 644 645 646 647 /// InsertUniqueBackedgeBlock - This method is called when the specified loop 648 /// has more than one backedge in it. If this occurs, revector all of these 649 /// backedges to target a new basic block and have that block branch to the loop 650 /// header. This ensures that loops have exactly one backedge. 651 /// 652 BasicBlock * 653 LoopSimplify::InsertUniqueBackedgeBlock(Loop *L, BasicBlock *Preheader) { 654 assert(L->getNumBackEdges() > 1 && "Must have > 1 backedge!"); 655 656 // Get information about the loop 657 BasicBlock *Header = L->getHeader(); 658 Function *F = Header->getParent(); 659 660 // Unique backedge insertion currently depends on having a preheader. 661 if (!Preheader) 662 return 0; 663 664 // The header is not a landing pad; preheader insertion should ensure this. 665 assert(!Header->isLandingPad() && "Can't insert backedge to landing pad"); 666 667 // Figure out which basic blocks contain back-edges to the loop header. 668 std::vector<BasicBlock*> BackedgeBlocks; 669 for (pred_iterator I = pred_begin(Header), E = pred_end(Header); I != E; ++I){ 670 BasicBlock *P = *I; 671 672 // Indirectbr edges cannot be split, so we must fail if we find one. 673 if (isa<IndirectBrInst>(P->getTerminator())) 674 return 0; 675 676 if (P != Preheader) BackedgeBlocks.push_back(P); 677 } 678 679 // Create and insert the new backedge block... 680 BasicBlock *BEBlock = BasicBlock::Create(Header->getContext(), 681 Header->getName()+".backedge", F); 682 BranchInst *BETerminator = BranchInst::Create(Header, BEBlock); 683 684 DEBUG(dbgs() << "LoopSimplify: Inserting unique backedge block " 685 << BEBlock->getName() << "\n"); 686 687 // Move the new backedge block to right after the last backedge block. 688 Function::iterator InsertPos = BackedgeBlocks.back(); ++InsertPos; 689 F->getBasicBlockList().splice(InsertPos, F->getBasicBlockList(), BEBlock); 690 691 // Now that the block has been inserted into the function, create PHI nodes in 692 // the backedge block which correspond to any PHI nodes in the header block. 693 for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) { 694 PHINode *PN = cast<PHINode>(I); 695 PHINode *NewPN = PHINode::Create(PN->getType(), BackedgeBlocks.size(), 696 PN->getName()+".be", BETerminator); 697 if (AA) AA->copyValue(PN, NewPN); 698 699 // Loop over the PHI node, moving all entries except the one for the 700 // preheader over to the new PHI node. 701 unsigned PreheaderIdx = ~0U; 702 bool HasUniqueIncomingValue = true; 703 Value *UniqueValue = 0; 704 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 705 BasicBlock *IBB = PN->getIncomingBlock(i); 706 Value *IV = PN->getIncomingValue(i); 707 if (IBB == Preheader) { 708 PreheaderIdx = i; 709 } else { 710 NewPN->addIncoming(IV, IBB); 711 if (HasUniqueIncomingValue) { 712 if (UniqueValue == 0) 713 UniqueValue = IV; 714 else if (UniqueValue != IV) 715 HasUniqueIncomingValue = false; 716 } 717 } 718 } 719 720 // Delete all of the incoming values from the old PN except the preheader's 721 assert(PreheaderIdx != ~0U && "PHI has no preheader entry??"); 722 if (PreheaderIdx != 0) { 723 PN->setIncomingValue(0, PN->getIncomingValue(PreheaderIdx)); 724 PN->setIncomingBlock(0, PN->getIncomingBlock(PreheaderIdx)); 725 } 726 // Nuke all entries except the zero'th. 727 for (unsigned i = 0, e = PN->getNumIncomingValues()-1; i != e; ++i) 728 PN->removeIncomingValue(e-i, false); 729 730 // Finally, add the newly constructed PHI node as the entry for the BEBlock. 731 PN->addIncoming(NewPN, BEBlock); 732 733 // As an optimization, if all incoming values in the new PhiNode (which is a 734 // subset of the incoming values of the old PHI node) have the same value, 735 // eliminate the PHI Node. 736 if (HasUniqueIncomingValue) { 737 NewPN->replaceAllUsesWith(UniqueValue); 738 if (AA) AA->deleteValue(NewPN); 739 BEBlock->getInstList().erase(NewPN); 740 } 741 } 742 743 // Now that all of the PHI nodes have been inserted and adjusted, modify the 744 // backedge blocks to just to the BEBlock instead of the header. 745 for (unsigned i = 0, e = BackedgeBlocks.size(); i != e; ++i) { 746 TerminatorInst *TI = BackedgeBlocks[i]->getTerminator(); 747 for (unsigned Op = 0, e = TI->getNumSuccessors(); Op != e; ++Op) 748 if (TI->getSuccessor(Op) == Header) 749 TI->setSuccessor(Op, BEBlock); 750 } 751 752 //===--- Update all analyses which we must preserve now -----------------===// 753 754 // Update Loop Information - we know that this block is now in the current 755 // loop and all parent loops. 756 L->addBasicBlockToLoop(BEBlock, LI->getBase()); 757 758 // Update dominator information 759 DT->splitBlock(BEBlock); 760 761 return BEBlock; 762 } 763 764 void LoopSimplify::verifyAnalysis() const { 765 // It used to be possible to just assert L->isLoopSimplifyForm(), however 766 // with the introduction of indirectbr, there are now cases where it's 767 // not possible to transform a loop as necessary. We can at least check 768 // that there is an indirectbr near any time there's trouble. 769 770 // Indirectbr can interfere with preheader and unique backedge insertion. 771 if (!L->getLoopPreheader() || !L->getLoopLatch()) { 772 bool HasIndBrPred = false; 773 for (pred_iterator PI = pred_begin(L->getHeader()), 774 PE = pred_end(L->getHeader()); PI != PE; ++PI) 775 if (isa<IndirectBrInst>((*PI)->getTerminator())) { 776 HasIndBrPred = true; 777 break; 778 } 779 assert(HasIndBrPred && 780 "LoopSimplify has no excuse for missing loop header info!"); 781 (void)HasIndBrPred; 782 } 783 784 // Indirectbr can interfere with exit block canonicalization. 785 if (!L->hasDedicatedExits()) { 786 bool HasIndBrExiting = false; 787 SmallVector<BasicBlock*, 8> ExitingBlocks; 788 L->getExitingBlocks(ExitingBlocks); 789 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 790 if (isa<IndirectBrInst>((ExitingBlocks[i])->getTerminator())) { 791 HasIndBrExiting = true; 792 break; 793 } 794 } 795 796 assert(HasIndBrExiting && 797 "LoopSimplify has no excuse for missing exit block info!"); 798 (void)HasIndBrExiting; 799 } 800 } 801